Quick take: This morning’s stories share a useful discipline: ask who bears the cost when a promising system scales. Better metabolic treatment, automated pricing, data-centre power demand and AI training all become public questions when their gains and burdens fall on different people.
A stronger obesity result comes with a higher treatment burden
On 30 September, Eli Lilly reported a 48-week Phase 2b trial in 367 adults with obesity or overweight and type 2 diabetes. Its experimental combination of eloralintide, an amylin-receptor agonist, and tirzepatide produced average weight loss of up to 23.3%, compared with 14.8% for tirzepatide alone; glycated haemoglobin also fell further. That is promising because two biological pathways may help people who need both weight and glucose control. But these are company-reported, mid-stage results, not evidence of approval or durable benefit. Gastrointestinal effects were common, and discontinuation reached 27% in one combination group during dose escalation. The practical lesson is not to chase a headline number. Patients can ask clinicians how trial populations, side-effect withdrawals and long-term maintenance compare with their own circumstances, while watching for peer-reviewed data and larger Phase 3 results. Independent report.
Possible Alkemata article: When More Weight Loss Also Means More Treatment Burden
A judge pauses New York’s ban on rent-setting software
A federal judge on 29 September temporarily blocked New York from enforcing a law aimed at rent-setting software while RealPage’s challenge proceeds. Judge Valerie Caproni called the question close but found the company marginally likely to show that the law’s broad restriction on software recommendations violates the First Amendment, including when recommendations rely on public information. The ruling is preliminary: it does not finally validate the software or erase separate antitrust scrutiny. The promising part is institutional rather than technological—courts can force lawmakers to define prohibited conduct precisely. The danger is that renters lose a protection during litigation even as shared data and common algorithms may help landlords move prices together. For individuals, the machinery can appear only as an unexplained renewal increase. Renters can preserve renewal notices and comparable listings, ask landlords what inputs shaped a rise, and support rules targeted at non-public data sharing and coordinated conduct rather than vague bans on “algorithms.”
Possible Alkemata article: When a Rent Recommendation Becomes Market Coordination
US regulators tell PJM to revisit who pays for data-centre growth
The materially new development is a federal order, not another electricity-demand forecast. On 29 September, the Federal Energy Regulatory Commission accepted the core of PJM’s proposed reliability backstop but delayed the procurement and directed revisions so costs are assigned more closely to the zones and customers driving rapid data-centre load growth. PJM, which coordinates power across 13 states and Washington, projects a capacity shortfall of roughly 6.8 gigawatts. A targeted backstop could bring generation or demand reductions online faster; the promise is reliability without quietly socialising every expense. The uncertainty is whether “cost causation” can be measured fairly when new infrastructure serves both large loads and the wider grid. Households may feel the answer in bills long before they see a new power plant. Customers and local officials can inspect utility rate filings, demand transparent load forecasts and ask which users pay for upgrades. FERC’s concurrence sets out the reasoning.
Possible Alkemata article: Who Pays When Data Centres Outgrow the Grid?
A US appeals court draws a narrow line around AI training and fair use
In an opinion issued on 29 September and unsealed the next day, the US Court of Appeals for the Third Circuit upheld Thomson Reuters’ copyright victory over Ross Intelligence. Ross had used Westlaw headnotes to build a competing legal-search system; the court found the headnotes sufficiently original and the use insufficiently transformative to qualify as fair use. This is the first US appellate ruling on fair use in AI training, but its reach is deliberately narrow: Ross was not a generative-AI system, and the judges distinguished broader disputes involving generative models. The promise is clearer evidence that technical processing does not dissolve ordinary questions about copying, licensing and market substitution. The danger is overreading one fact-specific dispute as a universal answer, chilling useful research or encouraging careless training elsewhere. Developers and organisations procuring AI can act now: keep source and licence records, document why each dataset is used, and test whether an output or product substitutes for the source market. Independent report.
Possible Alkemata article: What an AI Developer Must Know About Training-Data Provenance
